AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Training–Inference Consistent Segmented Execution for Long-Context LLMs

Xianpeng Shang, Jiang Li, Zehua Duo, Qianyi Cai, Xiangdong Su

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
align training with segmented inference execution instead of full-context training (ablation)segmented execution framework · LongBench-Eaverage score: 11.91 -> 24.17Improved
increase TBPTT depth from K=1 to K=2 (ablation)segmented execution framework · LongBench-Eaverage score: 24.17 -> 24.07No clear effect
use training-inference consistent segmented execution vs vanilla attentionLLaMA2-7B-32K · LongBench-E summarizationscore: 3.09 -> 12.10Improved
use segmented execution vs full attention on average LongBench-ELLaMA2-7B-32K · LongBench-Eaverage score: 23.13 -> 23.24No clear effect
use segmented execution vs full attention on average LongBench-ELLaMA2-7B-80K · LongBench-Eaverage score: 23.38 -> 24.17Improved
use segmented execution to cut prefill memoryLLaMA2-7B-80K · 32K prefillpeak memory (GB): 34.67 -> 19.06Improved
use segmented execution to cut prefill memoryLLaMA2-7B-32K · 16K prefillpeak memory (GB): 23.61 -> 18.56Improved
use segmented execution instead of vanilla attentionLLaMA2-7B-32K · 16K prefillTTFT (s): 1.62 -> 1.70Hurt
use segmented execution on RULER beyond training length (64K)LLaMA2-7B-32K · RULER CWE 64Kaccuracy: - -> 2.00Improved
use segmented execution on RULER CWE and FWE 4K-32KLLaMA2-7B-32K · RULER CWE/FWE avg 4K-32Kaccuracy: CWE 32.94 -> 46.39; FWE 41.33 -> 43.88Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy